PPG-GAN: An Adversarial Network to De-noise PPG Signals during Physical Activity

Abstract

Quality photoplethysmographic (PPG) signals are essential for accurate physiological assessment. However, the PPG acquisition process is often accompanied by spurious motion artefacts (MAs), especially during medium-high intensity physical activity. This study proposes a generative adversarial network (PPG-GAN) to create de-noised versions of measure PPG signals. The Adaptive Notch Filtration (ANF) algorithm, which enables the extraction of accurate heart rates (HR) and respiration rates (RR) from PPG signals, is used as the approximate reference signal to train the PPG-GAN. The generated PPG signals from test inputs provide a heart rate (HR) with a mean absolute error of 1.68 bpm for the IEEE-SPC dataset. A comparison with gold-standard HR and RR measurements, for our in-house dataset, show the errors in absolute value of less than 5%. The generated PPG signals, for the test clips, show a very strong correlation with their reference values, R ≈ 0.98. The results suggest that PPG-GAN could be a paradigm for MA-free PPG signal processing specifically for personal healthcare, even during high intensity activity.